A load forecasting method based on classified identification deep belief network is proposed to solve the problem ofslow convergence and large prediction error of traditional neural network in load forecasting. In the proposed method

the input historicalload data is normalized firstly. Then

the layered pretraining of thedeep believe machine is implemented by hierarchical unsupervisedgreedy pre training method. Thirdly

the pre training results areused as the initial value of the supervised learning training model.Deep belief network is composed of multilayer restricted boltzmannmachine

and the pre weight of the restricted boltzmann machineand the classification recognition mechanism is trained by means of contrastive divergence. The learning performance of the classified identification deep belief network can be improved by this way. Simulation result shows that the load forecasting algorithm based on the200 times of load training and temperature training

the algorithm has faster convergence speed and higher prediction accuracy than the selforganizing fuzzy neural network and the BP neural network.